John Egan

Are you a startup wanting to build a GenAI-based RAG app (production-grade) for customer support in less than a month? Here’s the step 👇 1) Start with your docs -> chunk & embed them 2) Store in a vector DB (Pinecone/Weaviate) 3) Retrieve the most relevant chunks -> feed into an LLM 4) Add citations + safety filters -> deploy as an assistant The best part... With AWS Bedrock’s serverless Knowledge Base, you can skip a lot of infra headaches. Managed RAG Pipeline and enterprise-grade security. Check the architecture below 👇 Perfect for startups that want to move fast + safe. Build lean, validate with real users, and scale confidently. ✨ That’s how you ship GenAI-powered customer support in weeks, not months. You have a chance to get funding from AWS to build your PoC if you have a DPIIT certification and a valid use case, as per their policy. DM me if you’d like to benefit from this, and I’ll connect you with the concerned person.

37

Really excited for Plaid’s Effects. AI is rapidly raising the bar for financial products. The next generation of experiences will be proactive, personalized, and much better at understanding what’s actually happening in a user’s financial life. That shift is what I’m most excited to talk about: how better data, models, and infrastructure are turning connected finance into intelligent finance. If you care about where fraud, payments, lending, and AI are headed next, join us at Effects. https://lnkd.in/gdYdAZaH

126

2 Comments

This session we did two weeks ago was a Masterclass in bringing highly informed CISOs to discuss pressing topics in security and resilience. Well run by Gaven Smith CB and Chris Martin. Particularly interesting to me was the discussion on enforcement and incentives for securing organizations--it was a shame regulators weren't in the room taking notes!

1

New Dave's Quick Hits is up. Three stories with the same flaw. 1️⃣ Vitalik Buterin on why smart people justify terrible decisions with brilliant-sounding arguments. His test: if a scam could use the exact same reasoning, the argument tells you nothing. "Responsible acceleration" fails this test. So does "inevitable transformation." 2️⃣ OpenAI led a $15M seed into Red Queen Bio to evaluate whether AI bio capabilities are matched by adequate defenses. The structure: OpenAI builds the capabilities, funds the evaluator, and the evaluator judges if OpenAI is safe. That's like letting the developer inspect their own building. 3️⃣ A method to test 100 samples of a father's DNA before making an embryo. Cost: $20,000 (same as IVF today). Result: your daughter moves from 50th percentile to 90th for traits genes control. The argument that follows: "It's irresponsible not to optimize your child." The pattern: arguments that sound sophisticated but rule nothing out. They work equally well for a legitimate company, a scam, or the opposite strategy. Vitalik's test: would this feel safe if your enemy ran it instead of someone you respect? Rules only hold when they're set before capital shows up. After that, revenue bends every line. Full analysis: https://lnkd.in/gkKk-W2F #AI #Biotech #VentureCapital #Innovation #vibecap #vibecapital

5

The XLSX format is not built for document extraction. Spreadsheets are not just glorified tables - what's just as important is the structure, formatting, relationships, and formulas that cascade across sheets. At Pulse, we trained a new spreadsheet encoder that teaches LLMs to understand large Excel files from the ground up. Not just cell contents, but the architecture. How conditional formatting triggers, where pivot tables pull their ranges, and why merged cells break everything. The result was reflected in our latest launch, Pulse Meridian: when our system processes a financial PDF, it doesn't just extract numbers, it reconstructs the spreadsheet logic: • Headers that span properly across complex table structures  • Formulas that reference the right cells automatically This novel approach is reflected inside the Pulse API - the result is being able to fundamentally speak Excel. Feel free to try it now, and email us with your most complex spreadsheets!

38

2 Comments

Parsing PDFs at scale with LLMs is cost prohibitive. Newer models (e.g. gemini 3) are good at reading pdfs, but you burn unnecessary vision tokens even when the page is text heavy. We’ve built in a “cost-optimizer” within LlamaParse that will dynamically route pages to fast/cheap parsing depending on its complexity.  Complex pages (e.g. those with tables/charts/diagrams) will still get routed to our VLM-enabled modes. This will let you save anywhere from 50-90% of parsing costs, at much higher accuracy compared to the comparable mode of feeding screenshots into VLMs. Check it out! https://lnkd.in/g9Wpqn7w

77

3 Comments

This is very true. The valuation trap is real. But I don’t think this is a YC story. It’s a market story. We’re seeing the exact same thing here in Israel. Seed rounds at 30–40M for companies with no product, no real validation, sometimes barely a prototype. Not second-time founders. First-timers. I’ve personally seen more than 20 rounds like this just last year. This isn’t about YC “pushing” companies. The system as a whole hasn’t adapted. The cost of building collapsed. The speed of iteration exploded. But pricing logic stayed anchored in 2021. When capital doesn’t recalibrate to reality, it doesn’t just create paper risk. It creates structural fragility. Good companies get boxed into impossible expectations. They are forced to grow into valuations instead of growing into product-market fit. That’s how promising businesses die early. The tragedy isn’t overvaluation. The tragedy is killing companies that could have become durable, profitable, meaningful businesses. This isn’t a YC issue. It’s a venture industry inertia issue. And the longer it takes to adjust, the more collateral damage we’ll keep seeing.

5

Interesting blog post from Airbnb that describes how it built an Embedding-Based Retrieval (EBR) system for home searches. The search relevancy problem is relevant for any consumer platform that must sift through a large field of candidate items when scoring search results, but it's idiosyncratically complex for Airbnb given 1) inconsistent query parameters on the part of a user (eg., group size can change from one trip to another given different trip purposes) and 2) flexible filters in query construction (eg., open-ended vs. fixed date ranges). Airbnb's EBR pipeline consists of 1) generating listing embeddings, 2) training a model on those embeddings as a first pass, and 3) running an Approximate Nearest Neighbors (ANN) search to surface the best K candidates. The embeddings were generated using contrastive learning on queries with positive and negative pairs of candidate listings. Since booking a home on Airbnb often involves a multi-step process (search, click on listing, view pictures, read reviews, etc.), Airbnb could generate the negative listings from those that the user interacted with but didn't ultimately book. This differs from another common approach in contrastive learning, which is to simply use random items for the negative pairs. With the listing embeddings available, Airbnb could implement a Two Towers network model by calculating user embeddings on the fly (listing embeddings are generated with a nightly processing job). Euclidean Distance was used for similarity assessment; while dot product similarity performed similarly, Euclidean distance produced more uniform cluster sizes. The blog hypothesizes that this is because Euclidean distance accounts for vector magnitude, which, in this case, includes useful information such as "historical counts" (it's not clear from the post what those counts pertain to). For final retrieval, Airbnb implemented inverted file index (IVF) search. Clusters are pre-processed, and IVF requires storing only the clusters' centroids and the listings' indices, resulting in lower latency than other search methods. The EBR system captures more context from the search query when retrieving candidate listings. Airbnb notes in the blog post that its EBR system resulted in "a statistically-significant gain in overall bookings" when A/B tested against the legacy system. Blog post linked below.

30

1 Comment

I’m gonna hold your hand as I say this: Stop using Claude Cowork. If you’re serious about getting good at AI workflows, it’s the wrong place to spend time. Many of my incredibly sharp friends are defaulting to Claude Cowork because they think it’s the best fit for non-developers. It’s not. They’re trading real power for a small amount of convenience. That's because Claude Cowork is optimized for completing tasks. Claude Code is optimized for building systems you can reuse, extend, and share. And most importantly, the most powerful new capabilities show up in developer environments first. Tool use, memory, agents that operate across systems. That all lands in Claude Code before it trickles down anywhere else. Anthropic is doing a great job with Claude Cowork, and it will keep getting better. But it’s always going to lag behind Claude Code in capability and control. There’s a learning curve with Claude Code, but it’s mostly psychological. You don’t need to be an engineer. You just need to be willing to be uncomfortable for a bit. We’re in a moment where new tools are shipping constantly, and the real constraint is your time. If you’re going to invest hours getting better at this, you want it to compound. For me, that’s Claude Code. If I’m wrong, I’m curious: What does Claude Cowork unlock for power users that Claude Code doesn’t?

364

121 Comments

Love it or hate it, being in the SF Bay Area is still one of the biggest hacks for First Builders to discover huge opportunities before anyone else. But what can you do if you’re not here? Cecilia for example moved here in 2001 and instantly joined Yahoo after feeling the palpable energy around early internet. And it’s been nonstop ever since as she later joined Amazon, Cruise, Replit, and her own breakout company GC AI. She credits a lot to the emerging tech density and optimism in the Bay Area. But she also mentions additional ways people can identify these opportunities early on: - Follow intuition around big waves - Follow your envy - Never lose curiosity - Watch the teen hackers - Talk to peers early and often (builds long term networks/insights)

21

3 Comments

8 months to $100M ARR. That’s not a growth story. That’s a platform shift. What Mukund Jha, Madhav Jha and the Emergent team have proven is simple: software creation just crossed a structural inflection point. The distance between idea and deployed product is collapsing — to natural language, across platforms, for anyone! Becoming one of the fastest-growing startups ever couldn’t have come at a better time. With India AI Impact Summit 2026 putting the country firmly on the global AI map, this moment feels symbolic — a world-class AI company, built from India, scaling at global speed. We at Together are proud to have backed Emergent from day one. This is what the early innings of a generational company look like. 🚀 #AgenticAI #NextBillionBuilders #IndiaAI

19

1 Comment

Why we killed a $5M ARR product to build Conversion. Okay, we didn't kill it. I'm being dramatic. But, we still decided to focus almost all of our new product development entirely on a new product with $0 in revenue. Why? Because when it comes to real product market fit that can lead to top decile growth, there are 4 boxes you have to check: 1/ The person with the pain – They feel the problem every day. 2/ The person with the budget – They can pay for it. 3/ The person with the authority – They can sign the deal. 4/ The person who will actually use it – They have the keys to deploy. If all four are the same person, you have a rocket ship. Our old product had two or three of these at best. It made money, but every deal felt like pushing a boulder uphill. SEO specialists or freelancers didn't have the budget nor authority to get it through. Demand-gen or marketing leaders didn't feel the pain or use it. With Conversion, all four map to the same person: the modern B2B marketing leader. They feel the pain. They own the budget. They can sign. And they can log in tomorrow and start running campaigns. We also learned something critical: urgent > important. A big TAM and a real problem do not matter if the customer does not feel urgency. Conversion solves problems that block revenue today, which makes deals move fast. The result? Bigger commitments, more sophisticated customers, better feedback loops, and faster revenue growth. Thank you to Gili Raanan for this gem of a framework!

131

8 Comments

My friend runs a DTC supplement brand. Clean greens powder. Was doing about $47K a month and wanted to scale faster. I said listen, you need AI automation. Full stack. They went all in. Claude API. n8n workflows. Built entire content engine scraping Reddit health threads and auto-generating blog posts. Complete revolution. Result ... Still doing $47K a month. But now their tools cost $1890/month and they spend 15 hours a week debugging webhooks. It's data-driven stagnation though. AI-powered loss with full tracking so they know exactly which automated system isn't working. Their CMO suggested they just... talk to their customers and improve the product taste. They said he doesn't understand AI led growth hacking.

24

ctac just dropped some solid results (see below) they connected underserved communities with research over 18 months and launched 20+ trials reaching 4,600+ participants (mckinsey helped with the project management) the headline stat everyone's sharing? "50% of trials happen in just 2% of zip codes" honestly that feels a bit misleading... of course academic medical centers concentrate trials, that's where the research infrastructure actually exists. the real question isn't about zip codes, it's about getting viable sites up and running in underserved areas what actually worked: site readiness assessments (no more "we'll figure it out later") mentorship between experienced sites and newcomers platforms that actually connect sponsors with trial-ready locations but honestly? building research infrastructure shouldn't feel revolutionary in 2025 at power we see the same thing... remove the friction and suddenly patients who've been invisible to the system start enrolling. turns out people want access to cutting-edge treatments regardless of their zip code (shocking, i know)the future of medicine literally depends on everyone being able to participate, not just people who live near academic medical centers what's your take - is geography still the biggest barrier or have we moved past that? 

7

AI ROI is being mispriced. Everyone is counting prompts, seats, token costs. The real bill is org design. - AI that saves minutes changes nothing - AI that removes handoffs changes headcount math - AI that works only with consultants is still services, not software We’re watching the same mistake as early SaaS. People are buying tools. The winners will buy labor redesign.

38

11 Comments

We, at Blitzscaling Ventures have invested in Ambiguous: AI agent teammates for business. Are they 1) AI agents helping my business, or 2) are they remote humans? We can't tell, it's Ambiguous! Today’s work tools assume either humans do the work, or AI replaces it. Ambiguous takes a different approach: AI coworkers that operate alongside humans in the same workspace, with real identities (each with Google accounts), shared context, and a full audit trail. Humans and AI collaborate on the same threads across Google Workspace, Slack, and more, making work feel less like automation and more like teamwork. The product is designed around product-led growth and virality (key to our investment thesis): AI coworkers naturally interact with other humans, spreading through shared threads, emails, and workflows. As teams add more coworkers and expand use cases, the platform compounds in value, moving from individual productivity to team operations and eventually enterprise-wide adoption. The founding team Ryan Waliany, Philip Lee, Dan Hsiao all worked together at a prior company Doorstead, brings a rare combination of deep technical skill and operating experience. We met the Ambiguous team through the a16z speedrun and were immediately impressed by their clarity of thought and ambition. We believe Ambiguous is defining a new category: not just AI assistants, but true AI coworkers that work like real teammates. We’re excited to partner with the team as they build the operating system for agentic work.

93

21 Comments

Incredibly excited to announce our acquisition of Runebook (acq. by Keycard), and to share more about the work we’re already doing with customers to deploy production-ready MCP agents and tools into production. When we met with Peter Cho and Matte Noble earlier this year at MCP Night, it was immediately clear they were thinking about the future of computing in a special way. Like us, they saw MCP as the fulcrum for bringing the advances of LLMs into the systems that power modern life, but also understood that it was too difficult to adopt and challenging to secure. So when the opportunity came to bring them into the fold, it was an easy decision. They’ve been working to make MCP easier to adopt, and with their experience building beloved developer ecosystems at Heroku and Sentry, they’ll expand the work we’re already doing to let developers build trusted agentic applications in minutes — not days or weeks. You can’t become agent-native without enabling your developers to build, your employees to adopt, and your security teams to govern with full observability and policy enforcement out of the box. We’re helping customers do exactly that today. More in the blog post, linked in the comments below 👇

19

2 Comments

We just did a big analysis of Clay's 10,000 monthly support tickets. If you're running a support team, here's what you need to know: A few weeks ago, George Dilthey (Head of Support), went to Karan Parekh with a problem, "Hey, the org is growing really fast. We're hiring all these support people, but they still feel underwater. Help me figure out what's going on." The obvious answer? Hire more people. But candidly, we hadn't really studied the data yet. The thing we needed to know first was, how are we getting through ticket volumes? When are they coming in? What time of the day? From what geography? Then once we had that data, we had to go back to first principles: What is the actual point of support? Of course, I want to give our customers an excellent experience. I want it to be world-class. I want it to solve their problem quickly and expediently. But how would you define expediently? Is it zero wait-times? 10 minutes? 60 minutes? In this video, Karan and I break down: - The bimodal curve we discovered -- tickets aren't flat, there's a pattern you can plan for - Why Thursday-Friday have half the volume of Monday-Wednesday (and what we do with that capacity) - How to staff for sub-1 hour response times without having people idle most of the day - Our weekend coverage strategy (that doesn't burn people out and eliminates the Monday morning queue) - What our support team actually does during low-volume periods If your support team feels underwater, watch below 👇 PS I don't know why the team chose a thumbnail where my eyes are closed 🙈

278

44 Comments